Arm starts selling the chip instead of the blueprint, and lands in its licensees' rack slot

A royalty is typically a small percentage of a chip's selling price ^[10]. Selling the chip is the whole price. That gap is the financial case here, and it shows up in the shape of the target: $15B of the $25B Arm wants annually is meant to come from silicon, leaving $10B to licensing ^[6]. Three fifths of the intended company is work Arm has never done, since it has never manufactured or sold chips at scale ^[1][11].

The timing deserves underlining. Meaningful AGI CPU revenue is not expected before the second half of 2026, the larger contribution arrives in fiscal 2028, and the production ramp starts in the back half of this year ^[7]. The committed order book through fiscal 2028 is roughly one seventh of the annual silicon figure Arm is aiming at for around 2031 ^[2]. Everything between those two numbers is a wager on x86 sockets changing hands, a market Intel and AMD have historically held and that Arm-based server parts from Amazon and Ampere have already been eroding on power efficiency ^[13].

Investors spent a moment unhappy about their architecture supplier becoming a rival to its own licensees, then pushed the stock up roughly 16 to 20 percent ^[8]. The awkwardness survives the rebound. Amazon's Graviton and Google's in-house Arm server processors are royalty-paying products that now sit where Arm wants to sell ^[9], and Qualcomm and MediaTek read the precedent whether or not they ever buy a server CPU ^[9].

The pitch to buyers is total cost rather than silicon: more than twice the performance per rack against x86, and up to $10B of capital expenditure avoided per gigawatt of capacity deployed ^[4]. That is Arm's own modelling, and it is benchmarked against x86 rather than against the Arm server parts hyperscalers already build for themselves, which is the comparison a Graviton operator would actually run. Underneath it sits a density claim: 136 Neoverse V3 cores in a 300-watt envelope is about 2.2 watts per core ^[2][3], aimed at agentic inference, where Arm says core requirements rise as much as fourfold ^[4].

There is one exposure the structure cannot hedge. Every discount Arm concedes to win a Meta-scale socket lands on the same income statement that still expects $10B a year from the companies it is bidding against ^[6][9]. Licensing revenue was insulated from who won; silicon revenue is not.